Skip to main content

A sentence-level visualization tool for attention in large language models

Project description

SAVIS: Sentence-Level Attention Visualization for Large Language Models

SAVIS (Sentence-level Attention VISualization) is a Python library for visualizing inter-sentence attention in large language models (LLMs), specifically designed for public relations (PR) analysis. This tool enhances the interpretability of LLMs by providing an intuitive visualization of how attention is distributed across sentences in generated text.

SAVIS analyzing reviews

Features

  • Inter-sentence attention calculation
  • Interactive visualization of attention patterns
  • Support for various LLMs through Hugging Face's transformers library
  • Customizable for PR-specific analysis tasks with Jupyter Notebook

Installation

pip install savis

Quick Start

from savis import TextGenerator, ISA, ISAVisualization

# Initialize the text generator with your chosen model
generator = TextGenerator("Model name")

# Generate text and get attention data
input_text = "Your input prompt here"
generated_text, attentions, tokenizer, input_ids, outputs = generator.generate_text(input_text)

# Calculate inter-sentence attention
isa = ISA(outputs.sequences[0], attentions, tokenizer)

# Visualize the attention patterns
vis = ISAVisualization(isa.sentence_attention, isa.sentences)
vis.visualize_sentence_attention()

Key Components

  1. TextGenerator: Interfaces with the LLM to generate text and extract attention information.
  2. Attention: Manages the underlying LLM and provides methods for obtaining attention data from the model.
  3. ISA (Inter-Sentence Attention): Processes raw attention data to compute attention between sentences.
  4. ISAVisualization: Creates interactive visualizations of the computed inter-sentence attention.

These components work together to provide a comprehensive pipeline from text generation to attention visualization:

  • TextGenerator uses the LLM to generate text based on input prompts.
  • Attention handles the interaction with the LLM, extracting detailed attention information.
  • ISA takes the raw attention data and computes meaningful inter-sentence attention scores.
  • ISAVisualization takes these scores and creates interactive visualizations.

License

SAVIS is released under the MIT License. See the LICENSE file for more details.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

savis-0.3.6.tar.gz (4.9 MB view hashes)

Uploaded Source

Built Distribution

savis-0.3.6-py3-none-any.whl (9.6 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page